Incorporating Convolution Designs into Visual Transformers

Incorporating Convolution Designs into Visual Transformers

**动机:**然而,纯变压器架构通常需要大量的训练数据或额外的监督,以获得与卷积神经网络(CNNs)相当的性能。为了克服这些限制,我们分析了直接从NLP中借用变压器架构时的潜在缺陷。在此基础上,==我们提出了一种新的卷积增强图像变压器(CeiT),==它结合了cnn在提取低水平特征、增强局部性和变压器在建立远程依赖关系方面的优势。对原变压器进行了三个修改:

1)我们设计了一个图像到令牌(I2T)模块,而不是直接的标记化,该模块从生成的低级特征中提取补丁

2)每个编码器块中的前馈网络被替换为局部增强的前馈(LeFF)层,促进空间维度中相邻标记之间的相关性;

3)在利用多层次表示的变压器顶部附加了一个分层类标记注意(LCA)。

引言:

回顾卷积,其主要特征是**平移不变性和局部性**[22,31]。平移不变性与权值共享机制有关,该机制可以捕获视觉任务[23]中的几何和拓扑信息。对于局部性,在视觉任务[11,26,9]中,一个常见的假设是相邻的像素往往是相关的。然而,纯变压器架构并没有充分利用这些先前在图像中存在的偏差。首先,ViT对大小为16×16或32×32的原始输入图像中的补丁进行直接标记化。很难提取出在图像中形成一些基本结构的低级特征(如角落和边缘)。其次,自我注意模块集中于建立令牌之间的长期依赖关系,而忽略了空间维度中的局部性。

为了解决这些问题,我们设计了一种卷积增强图像变压器(CeiT),结合cnn在提取低级特征方面的优势、增强局部性以及变压器在关联远程依赖方面的优势。与普通ViT相比,做了三个修改。为了解决第一个问题,

我们设计了一个图像到标记(I2T)模块,而不是从原始输入图像中进行直接的标记化,该模块从生成的低级特征中提取补丁,其中补丁的尺寸更小,然后被平化为一系列标记。由于设计良好的结构,I2T模块不会引入更多的计算成本。

为了解决第二个问题,每个编码器块中的前馈网络被替换为局部增强的前馈(LeFF)层,该层促进了空间维度中相邻令牌之间的相关性。

为了利用自我注意的能力,在变压器的顶部附加了一个分层的类标记注意(LCA),利用多层表示来改进最终的表示。综上所述,我们的贡献如下:

  1. 我们设计了一种新的视觉变压器架构,即卷积增强图像变压器(CeiT)。它结合了卷积神经网络在提取低水平特征方面的优势,加强了局部性,以及变压器在建立远程依赖关系方面的优势。
  2. 在ImageNet和7个下游任务上的实验结果表明,与之前的ransformer 相比,CeiT的有效性和泛化能力学生和最先进的网络神经网络,不需要大量的培训数据和额外的CNN教师。例如,使用与ResNet- 50相似的模型大小,CeiT-S在ImageNet上获得了82.0%的前1位精度。当细化到384×384的分辨率时,结果提高到83.3%。
  3. CeiT模型比纯变压器模型具有更好的收敛性,训练迭代次数少3次×,显著降低了训练成本

3. Methodology

我们的CeiT是基于ViT而设计的。首先,我们将在第3.1节中简要概述了ViT的基本组件。接下来,我们将介绍三个包含卷积设计和有利于视觉变压器的修改,包括3.2节中的图像到令牌(I2T)模块、3.3节中的本地增强的FeedForwad(LeFF)模块和3.4节中的分层类令牌注意(LCA)模块。最后,我们在第3.5节中分析了这些所建议的模块的计算复杂度。

3.1.我们首先重新审视ViT中的基本组件,包括标记化、编码器块、多头自注意(MSA)层和前馈网络(FFN)层。

Tokenization.

标准的变压器[37]接收一系列的令牌嵌入作为输入。为了处理2D图像,ViT将图像 x ∈ R H × W × 3 x∈R_{H×W×3} xRH×W×3重构为扁平的2D补丁序列 x p ∈ R N × ( P 2 ⋅ 3 ) x_p∈R^{N×(P^2·3)} xpRN×P23),其中(H,W)为原始图像的分辨率,3为RGB图像的通道数,(P,P)为每个图像补丁的分辨率, N = H W / P 2 N = HW/P^2 N=HW/P2为得到的补丁数,也作为变压器的有效输入序列长度。这些补丁被扁平并映射到大小为C的潜在嵌入,然后在序列中添加一个额外的类标记作为图像表示,从而输入大小为 x t ∈ R ( N + 1 ) × C x_t∈R^{(N+1)×C} xtRN+1×C的序列.

在实践中,ViT将补丁大小为16×16或32×32分割每个图像。但是,对具有大补丁的输入图像的直接标记化可能有两个限制: 1)很难捕获图像中的低级信息(如边和角);2)大内核的参数化过多,往往难以优化,因此需要更多的训练样本或训练迭代。

编码器块。ViT是由一系列堆叠的编码器组成的。每个编码器都有MSA和FFN的两个子层。在每个子层周围使用一个残差连接[12],然后是层归一化(LN)[1]。每个编码器的输出为:

与cnn在每个阶段开始时进行降采样不同,在不同的编码器块中,标记的长度没有减少。有效的感受域不能有效地扩展,这可能会影响视觉变压器的优化效率。

MSA.

就是标准的注意力机制!

FFN.

FFN执行点级操作,并分别应用于每个令牌。它由两个线性变换组成,中间有非线性激活:

而σ(·)是ViT中GELU [13]的非线性激活。

作为MSA模块的补充,FFN模块对每个令牌进行维数展开/降简和非线性变换,从而提高了令牌的表示能力。然而,在视觉中很重要的标记之间的空间关系并没有被考虑。这导致原始的ViT需要大量的训练数据来学习这些归纳偏差

3.2.具有低水平特征的图像到令牌

为了解决上述标记化中的问题,我们提出了一个简单而有效的模块,名为图像-标记(I2T),该模块从特征映射中提取补丁,而不是从原始输入图像中提取补丁。如图2所示,I2T模块是一个由卷积层和最大池化层组成。消融研究还表明,在卷积层之后的批层有利于训练过程。它的可记号为:

其中,x‘∈R H S×WS×D,S为原始输入图像的步幅,D为富集通道数。然后将学习到的特征映射提取为空间维度上的斑块序列。为了保持生成的令牌数量与ViT保持一致,我们将补丁的分辨率缩小为 ( P / S , P / S ) (P/S,P/S) P/SP/S。在实践中,我们设置了S = 4。

I2T充分利用了cnn在提取低级特征方面的优势,通过缩小补丁大小,降低了嵌入的训练难度。这也不同于ViT中提出的变压器类型变压器,使用常规ResNet-50从最后两个阶段提取高级特征。我们的I2T要轻得多。

3.3. Locally-Enhanced Feed-Forward Network

为了结合cnn提取局部信息的优势和变压器建立随机依赖关系的能力,我们提出了一个局部增强的前馈网络(LeFF)层。在每个编码器块中,我们保持MSA模块不变,以保持捕获令牌之间的全局相似性的能力。相反,原来的前馈网络层被LeFF所取代。其结构如图3所示。

其结构如图3所示。一个LeFF模块执行以下步骤。

首先,给定由前一个MSA模块生成的标记 x h t ∈ R ( N + 1 ) × C x_h^t∈R^{(N+1)×C} xhtRN+1×C,我们将它们分成补丁标记 x h p ∈ R N × C x_h^p∈R^{N×C} xhpRN×C和相应的类标记 x h c ∈ R C x_h^c∈R^C xhcRC。采用线性投影将补丁标记的嵌入扩展到更高维的 x l 1 p ∈ R N × ( e × C ) x_{l1 }^p∈R^{N×(e×C}) xl1pRN×e×C,其中e是扩展率。

其次,根据相对于原始图像的位置,将补丁标记恢复到空间维度上的 x p s ∈ R √ N × √ N × ( e × C ) x^s_ p∈R^{√N×√N×(e×C)} xpsRN×N×e×C的“图像”。

第三,我们对这些恢复的补丁标记的核大小为k进行深度卷积,增强与相邻的 k 2 − 1 k^2−1 k21标记的表示相关性,得到 x p d ∈ R √ N × √ N × ( e × C ) x^d_p∈R^{√N×√N×(e×C)} xpdRN×N×e×C

第四,这些贴片标记被扁平为 x p f ∈ R N × ( e × C ) x^f_p∈R^{N×(e×C)} xpfRN×e×C)序列。最后,用 x p l 2 ∈ R N × C x^{l_2 }_p∈R^{N×C} xpl2RN×C将补丁标记投影到初始维度,并与类标记连接,得到 x t + 1 h ∈ R ( N + 1 ) × C x^h_ {t +1}∈R^{(N+1)×C} xt+1hRN+1×C。在每一个线性投影和深度卷积之后,添加一个批数和一个GELU。这些程序可记录如下:

image-20230421090529041

3.4.分层分类标记注意(Layer-wise Class-Token Attention)

在cnn中,随着网络的加深,特征图的感受野增加。在ViT中也发现了类似的观察结果,其“注意距离”随着深度的增加而增加。因此,特征表示在不同的层上将是不同的。为了集成跨不同层的信息,我们设计了一个分层的类标记注意(LCA)模块。与标准的ViT以最后一个l层的类令牌 x c ( L ) x^{ (L)}_ c xc(L)作为最终的表示不同,LCA关注不同层的类令牌。

如图4所示,LCA得到一个类标记序列作为输入,可以表示为 X c = [ x ( 1 ) c 、 ⋅ ⋅ ⋅ 、 x ( l ) c 、 ⋅ ⋅ ⋅ 、 x ( L ) c ] , X_c = [x_{(1)}^c、···、x_{(l)}^c、···、x_{(L)}^c ], Xc=[x(1)c⋅⋅⋅x(l)c⋅⋅⋅x(L)c]其中l表示图层的深度。LCA遵循标准的变压器块,其中包含一个MSA和一个FFN层。与最初的MSA计算任意两个标记(O(n 2))之间相似性的相似性不同,LCA==只计算最后一层中的类标记与其他层中的其他类标记(O (n))之间的相关性,==其中n表示标记的数量。对应的 x ( L ) c x_{(L)}^c x(L)c值通过注意与他人聚合。然后将聚合的值发送到一个FFN层中,从而得到最终的表示形式.

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from functools import partial

from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.registry import register_model
from timm.models.vision_transformer import default_cfgs, _cfg


__all__ = [
    'ceit_tiny_patch16_224', 'ceit_small_patch16_224', 'ceit_base_patch16_224',
    'ceit_tiny_patch16_384', 'ceit_small_patch16_384',
]

#Image-to-Tokens with Low-level Features
class Image2Tokens(nn.Module):
    def __init__(self, in_chans=3, out_chans=64, kernel_size=7, stride=2):
        super(Image2Tokens, self).__init__()
        self.conv = nn.Conv2d(in_chans, out_chans, kernel_size=kernel_size, stride=stride,
                              padding=kernel_size // 2, bias=False)
        self.bn = nn.BatchNorm2d(out_chans)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.maxpool(x)
        return x


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class LocallyEnhancedFeedForward(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.,
                 kernel_size=3, with_bn=True):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        # pointwise
        self.conv1 = nn.Conv2d(in_features, hidden_features, kernel_size=1, stride=1, padding=0)
        # depthwise
        self.conv2 = nn.Conv2d(
            hidden_features, hidden_features, kernel_size=kernel_size, stride=1,
            padding=(kernel_size - 1) // 2, groups=hidden_features
        )
        # pointwise
        self.conv3 = nn.Conv2d(hidden_features, out_features, kernel_size=1, stride=1, padding=0)
        self.act = act_layer()
        # self.drop = nn.Dropout(drop)

        self.with_bn = with_bn
        if self.with_bn:
            self.bn1 = nn.BatchNorm2d(hidden_features)
            self.bn2 = nn.BatchNorm2d(hidden_features)
            self.bn3 = nn.BatchNorm2d(out_features)

    def forward(self, x):
        b, n, k = x.size()#输入是一个序列
        cls_token, tokens = torch.split(x, [1, n - 1], dim=1)#第一维度是类别token
        x = tokens.reshape(b, int(math.sqrt(n - 1)), int(math.sqrt(n - 1)), k).permute(0, 3, 1, 2)#将tokenreshape到图像形状
        #无论是不是使用bn,最后都是对图像token进行处理
        if self.with_bn:
            x = self.conv1(x)
            x = self.bn1(x)
            x = self.act(x)
            x = self.conv2(x)
            x = self.bn2(x)
            x = self.act(x)
            x = self.conv3(x)
            x = self.bn3(x)
        else:
            x = self.conv1(x)
            x = self.act(x)
            x = self.conv2(x)
            x = self.act(x)
            x = self.conv3(x)
#再次将token变为序列,与类别token拼接
        tokens = x.flatten(2).permute(0, 2, 1)
        out = torch.cat((cls_token, tokens), dim=1)
        return out


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
        self.scale = qk_scale or head_dim ** -0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.attention_map = None

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        # self.attention_map = attn
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class AttentionLCA(Attention):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super(AttentionLCA, self).__init__(dim, num_heads, qkv_bias, qk_scale, attn_drop, proj_drop)
        self.dim = dim
        self.qkv_bias = qkv_bias
        
    def forward(self, x):

        q_weight = self.qkv.weight[:self.dim, :]
        q_bias = None if not self.qkv_bias else self.qkv.bias[:self.dim]
        kv_weight = self.qkv.weight[self.dim:, :]
        kv_bias = None if not self.qkv_bias else self.qkv.bias[self.dim:]
        
        B, N, C = x.shape
        _, last_token = torch.split(x, [N-1, 1], dim=1)
        
        q = F.linear(last_token, q_weight, q_bias)\
             .reshape(B, 1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        kv = F.linear(x, kv_weight, kv_bias)\
              .reshape(B, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        # self.attention_map = attn
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, 1, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Block(nn.Module):

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, kernel_size=3, with_bn=True, 
                 feedforward_type='leff'):
        super().__init__()
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.norm1 = norm_layer(dim)
        self.feedforward_type = feedforward_type

        if feedforward_type == 'leff':
            self.attn = Attention(
                dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
            self.leff = LocallyEnhancedFeedForward(
                in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop,
                kernel_size=kernel_size, with_bn=with_bn,
            )
        else:  # LCA
            self.attn = AttentionLCA(
                dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
            self.feedforward = Mlp(
                in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop
            )

    def forward(self, x):
        if self.feedforward_type == 'leff':
            x = x + self.drop_path(self.attn(self.norm1(x)))
            x = x + self.drop_path(self.leff(self.norm2(x)))
            return x, x[:, 0]
        else:  # LCA
            _, last_token = torch.split(x, [x.size(1)-1, 1], dim=1)
            x = last_token + self.drop_path(self.attn(self.norm1(x)))
            x = x + self.drop_path(self.feedforward(self.norm2(x)))
            return x


class HybridEmbed(nn.Module):
    """ CNN Feature Map Embedding
    Extract feature map from CNN, flatten, project to embedding dim.
    """
    def __init__(self, backbone, img_size=224, patch_size=16, feature_size=None, in_chans=3, embed_dim=768):
        super().__init__()
        assert isinstance(backbone, nn.Module)
        img_size = to_2tuple(img_size)
        self.img_size = img_size
        self.backbone = backbone
        if feature_size is None:
            with torch.no_grad():
                # FIXME this is hacky, but most reliable way of determining the exact dim of the output feature
                # map for all networks, the feature metadata has reliable channel and stride info, but using
                # stride to calc feature dim requires info about padding of each stage that isn't captured.
                training = backbone.training
                if training:
                    backbone.eval()
                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))
                if isinstance(o, (list, tuple)):
                    o = o[-1]  # last feature if backbone outputs list/tuple of features
                feature_size = o.shape[-2:]
                feature_dim = o.shape[1]
                backbone.train(training)
        else:
            feature_size = to_2tuple(feature_size)
            feature_dim = self.backbone.feature_info.channels()[-1]
        print('feature_size is {}, feature_dim is {}, patch_size is {}'.format(
            feature_size, feature_dim, patch_size
        ))
        self.num_patches = (feature_size[0] // patch_size) * (feature_size[1] // patch_size)
        self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x):
        x = self.backbone(x)
        if isinstance(x, (list, tuple)):
            x = x[-1]  # last feature if backbone outputs list/tuple of features
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x


class CeIT(nn.Module):
    def __init__(self,
                 img_size=224,
                 patch_size=16,
                 in_chans=3,
                 num_classes=1000,
                 embed_dim=768,
                 depth=12,
                 num_heads=12,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 drop_rate=0.,
                 attn_drop_rate=0.,
                 drop_path_rate=0.,
                 hybrid_backbone=None,
                 norm_layer=nn.LayerNorm,
                 leff_local_size=3,
                 leff_with_bn=True):
        """
        args:
            - img_size (:obj:`int`): input image size
            - patch_size (:obj:`int`): patch size
            - in_chans (:obj:`int`): input channels
            - num_classes (:obj:`int`): number of classes
            - embed_dim (:obj:`int`): embedding dimensions for tokens
            - depth (:obj:`int`): depth of encoder
            - num_heads (:obj:`int`): number of heads in multi-head self-attention
            - mlp_ratio (:obj:`float`): expand ratio in feedforward
            - qkv_bias (:obj:`bool`): whether to add bias for mlp of qkv
            - qk_scale (:obj:`float`): scale ratio for qk, default is head_dim ** -0.5
            - drop_rate (:obj:`float`): dropout rate in feedforward module after linear operation
                and projection drop rate in attention
            - attn_drop_rate (:obj:`float`): dropout rate for attention
            - drop_path_rate (:obj:`float`): drop_path rate after attention
            - hybrid_backbone (:obj:`nn.Module`): backbone e.g. resnet
            - norm_layer (:obj:`nn.Module`): normalization type
            - leff_local_size (:obj:`int`): kernel size in LocallyEnhancedFeedForward
            - leff_with_bn (:obj:`bool`): whether add bn in LocallyEnhancedFeedForward
        """
        super().__init__()
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models

        self.i2t = HybridEmbed(
            hybrid_backbone, img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
        num_patches = self.i2t.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
                kernel_size=leff_local_size, with_bn=leff_with_bn)
            for i in range(depth)])

        # without droppath
        self.lca = Block(
            dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=0., norm_layer=norm_layer,
            feedforward_type = 'lca'
        )
        self.pos_layer_embed = nn.Parameter(torch.zeros(1, depth, embed_dim))

        self.norm = norm_layer(embed_dim)

        # NOTE as per official impl, we could have a pre-logits representation dense layer + tanh here
        # self.repr = nn.Linear(embed_dim, representation_size)
        # self.repr_act = nn.Tanh()

        # Classifier head
        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()

        trunc_normal_(self.pos_embed, std=.02)
        trunc_normal_(self.cls_token, std=.02)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes, global_pool=''):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()

    def forward_features(self, x):
        B = x.shape[0]
        x = self.i2t(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
        x = torch.cat((cls_tokens, x), dim=1)
        x = x + self.pos_embed
        x = self.pos_drop(x)

        cls_token_list = []
        for blk in self.blocks:
            x, curr_cls_token = blk(x)
            cls_token_list.append(curr_cls_token)

        all_cls_token = torch.stack(cls_token_list, dim=1)  # B*D*K
        all_cls_token = all_cls_token + self.pos_layer_embed
        # attention over cls tokens
        last_cls_token = self.lca(all_cls_token)
        last_cls_token = self.norm(last_cls_token)

        return last_cls_token.view(B, -1)

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x


@register_model
def ceit_tiny_patch16_224(pretrained=False, **kwargs):
    """
    convolutional + pooling stem
    local enhanced feedforward
    attention over cls_tokens
    """
    i2t = Image2Tokens()
    model = CeIT(
        hybrid_backbone=i2t,
        patch_size=4, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    return model


@register_model
def ceit_small_patch16_224(pretrained=False, **kwargs):
    """
    convolutional + pooling stem
    local enhanced feedforward
    attention over cls_tokens
    """
    i2t = Image2Tokens()
    model = CeIT(
        hybrid_backbone=i2t,
        patch_size=4, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    return model


@register_model
def ceit_base_patch16_224(pretrained=False, **kwargs):
    """
    convolutional + pooling stem
    local enhanced feedforward
    attention over cls_tokens
    """
    i2t = Image2Tokens()
    model = CeIT(
        hybrid_backbone=i2t,
        patch_size=4, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    return model


@register_model
def ceit_tiny_patch16_384(pretrained=False, **kwargs):
    """
    convolutional + pooling stem
    local enhanced feedforward
    attention over cls_tokens
    """
    i2t = Image2Tokens()
    model = CeIT(
        hybrid_backbone=i2t, img_size=384,
        patch_size=4, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    return model


@register_model
def ceit_small_patch16_384(pretrained=False, **kwargs):
    """
    convolutional + pooling stem
    local enhanced feedforward
    attention over cls_tokens
    """
    i2t = Image2Tokens()
    model = CeIT(
        hybrid_backbone=i2t, img_size=384,
        patch_size=4, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    return model
  • 0
    点赞
  • 1
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值